Article

Why doubling food did not double some AI calorie estimates

For people using AI food logs, a new study separates errors in portion scaling from errors in food values—and shows what to check.

Editorial illustration for Why doubling food did not double some AI calorie estimates: a document represents the research briefing. Not documentary evidence.

The food was named and its weight was supplied in grams. Still, four of seven AI models increased their calorie estimates much less than the weight increased, according to a preprint submitted on 19 November 2026. For people using or building food logs, the finding is a reminder that supplying an exact weight does not guarantee the software will apply it consistently.

There was no blurry photograph to blame. The useful distinction is between food matching, calorie density and multiplication: three jobs that can disappear inside one plausible-looking total.

How much should the estimate change when the weight doubles?

Researchers Donato Riccio, Fabrizio Maturo and Iulia Brezeanu tested seven open-weight language models—models whose trained weights are publicly available—on 168 foods from a US Department of Agriculture database. Each food appeared in English at nine stated weights, ranging from a quarter to four times its database default amount. That default is a survey fallback when no quantity is reported, not a recommended serving. The experiment used text throughout.

For the same food at an unchanged calorie density, twice the weight should mean twice the calories. Here, calories means the kilocalories, or kcal, used in nutrition labelling.

For Mistral-Small-3.2-24B, the researchers reported a scaling coefficient of 0.709. Translating that fitted relationship into a doubling gives 2^0.709 ≈ 1.635: about 63% more estimated calories, rather than 100% more. This calculation summarises the authors’ fitted relationship across foods and weights; it is not an observed result for a particular doubled meal.

The problem could hide at an ordinary portion. For the four models with large scaling errors, the authors’ across-food summaries were only 2–10% above the reference at the default amount. At a quarter of that weight, they were 47–130% too high; at four times it, 20–34% too low. These are summary ratios across foods, not error ranges for every item. But this was not universal: the other three models scaled close to proportionally under the main prompt.

The implication for evaluation: checking one believable portion can miss an error that changes with portion size. The separate question of chatbot accuracy versus user satisfaction makes a related distinction—an answer’s appeal and its correctness need different checks.

A calculator fixes scaling, not the food value

Calorie density is simply calories per unit of food, commonly expressed per 100 grams. USDA’s documented portion calculation is ordinary arithmetic:

Calories in the portion = calories per 100 g × portion weight in grams ÷ 100.

For a hypothetical food with 200 kcal per 100 g, a 75 g portion works out to 150 kcal. A 150 g portion works out to 300 kcal. These are invented round numbers to illustrate the calculation, not measured food values or outputs from the study.

The researchers tried separating those jobs. They asked each model for a food’s calorie density, held that value fixed, then multiplied it by the weight in code. This was not a database lookup: the density still came from the model. The authors report that this two-step approach improved accuracy at the smallest and largest portions for all four models with large scaling errors.

But the counterexample matters. For Qwen3.5-35B-A3B, which already scaled nearly proportionally, the two-step estimate was less accurate than the direct answer at all three comparison weights shown in supplementary Table S9. Correct multiplication did not rescue an imperfect density estimate.

That is the distinction a doubling check cannot settle. A wrong calorie density can produce totals that double perfectly. Passing the arithmetic check tells you the quantity was scaled consistently, not that the right food or calorie value was selected.

Could a different prompt help?

Yes, in these experiments. In a separate three-weight test with different answer-generation settings, asking for step-by-step reasoning improved scaling in all four badly affected models and brought three close to proportionality. Providing the reference calorie density inside the prompt was another intervention; it also brought three of the four close to proportional scaling.

These are different interventions, not one universal fix. They show that the behaviour depended on how the task was posed, rather than establishing an unchangeable inability to calculate.

The authors interpret the compressed estimates as being pulled toward values for a typical serving. That is a proposed explanation of the pattern, not direct proof of what a model remembered or a complete account of its internal calculation.

What the study leaves unanswered

This was an English-language, single-database experiment with one named food per query. It did not evaluate commercial food-logging apps, photographs, freely described multi-part meals, other nutrient estimates or health outcomes. Results apply to the tested model versions, prompts and settings, not automatically to every product using a related model.

Earlier work such as NutriBench studied carbohydrate estimates from meal descriptions. It supplies context for the broader nutrition-estimation problem, not a replication of this calorie-scaling experiment. Better numerical answers and better health are also separate outcomes, as explained in what AI health-coaching evidence actually measures.

The authors disclose relevant business ties: Iulia Brezeanu is an owner of Geometry Lab, which operates the AI food-logging app Macropass. Donato Riccio collaborates with the company on research and reports no financial relationship. They state that neither the company nor the app funded the study or had a role in its design, conduct, analysis or reporting.

The findings remain author-reported preprint results. The paper says the code and model responses are in a repository kept private until article publication; its linked repository was not publicly available on 20 November.

What to check in an AI food log

For a weighed entry, look for the matched food and preparation, the calorie value per 100 g and the calculation connecting that value to the grams entered. Raw versus cooked descriptions and edible-weight conventions matter to the food match; USDA’s documentation distinguishes these inputs.

For product teams, the design implication is to keep the selected food and its density fixed when only the grams change, then calculate the new total from those stored inputs. Asking the model for a fresh density each time can reintroduce variation. Evaluation should check both agreement with the appropriate reference and proportional changes across weights.

For readers, doubling the stated weight of the same food should double the calculated calories, apart from rounding. That is a useful consistency check, not a certificate of nutritional accuracy. The kitchen scale supplies the grams. The food log still has to show what it did with them.